Multiple kernel learning techniques for ligand based virtual screening

There are various methods that have come into existence for similarity calculation between two different objects of similar domain that can be feature vectors, such as kernels. Multiple kernel learning methods show considerable level of prediction accuracy as compared to single kernels in existing systems. This paper primarily focuses on binary classification problems in cheminformatics specifically for ligand based drug discovery. The work finds an improvement of cross validation accuracy and reduced complexity as compared to earlier work in standard datasets like Predictive Toxicology and Mutagenicity. For 7-fold cross validation, the highest accuracy was of 67.375%.